預測直腸癌病患的復發危險因子

背景: 根據際癌症研究機構(International Agency for Research on Cancer, IARC)統計至2020年發生直腸癌病例數達到732,210例(佔3.8%)，成為全球第七名最常見的癌症。儘管癌症治療技術取得進展，直腸癌的預後得到了顯著改善，但復發仍是直腸癌患者的重大問題。本研究目的為針對直腸癌病患發生復發的危險因子進行分析，應用不同機器學習模式預測能力與臨床應用進行比較，後續提供臨床風險管理。方法: 臨床數據來自台灣的六所醫院癌症登記中心的直腸癌數據集，資料蒐集期間為2011年至2020年，共計3,403筆有效數據，其中包含復發749筆。根據文獻探與臨床醫師討論後決定16預測變數作為直腸癌復發的候選風險因子。使用預測模式包括：CART (Classification And Regression Tree)、RF (Random Forest)、C4.5 (C4.5 Decision tree)與C5.0 (C5.0 Decision tree)。本研究以SMOTE (Synthetic Minority Oversampling Technique)取樣方法解決樣本類別不均衡的問題。評估績效指標包括分類準確率(Accuracy)、敏感度(Sensitivity)、特異度(Specificity)、F度量評分(F1 score)、Kappa、Matthews相關係數 (Matthew's Correlation Coefficient, MCC)和曲線下面積(Area Under the Curve, AUC)來進行效能評估。結果: 在整體直腸癌發生復發分析中，重要危險因子特徵選取結果排序前三名為化療、整併期別、癌胚抗原CEA (Carcinoembryonic antigen)檢驗值；而有化療發生復發分析中，重要危險因子特徵選取結果排序前三名為癌胚抗原CEA檢驗值、整併期別、原發部位手術邊緣。研究結果顯示，C5.0與C4.5可以產生最佳分類及最具希望的結果來預測復發，並提供預測變數重要性的指標。在預測模式結果中，整體直腸癌發生復發分析以C5.0 (0.8635)的AUC值為最高，而有化療發生復發分析以C4.5 (0.8673)的AUC值為最高。決策曲線分析結果顯示，整體直腸癌發生復發分析以C5.0有最高的預測臨床效益，而有化療發生復發分析以CART和C4.5預測臨床效益表現較佳。在重分類淨改善效益指數分析結果顯示：只有整體直腸癌發生復發分析以C5.0有顯著預測效果。討論與結論: 隨著直腸癌化療治療日益複雜，復發的發生可能對長期生存構成威脅。識別增加復發風險的治療相關危險因子是癌症倖存者監測和管理方面日益關注的問題。這些發現可能有助於為高危直腸癌倖存者在追蹤期間制定有效的預防和監測計劃。在未來，可以採取更先進的方法來解決這個階層的不平衡問題。這是本研究的局限性和未來的研究方向之一。

關鍵字

直腸癌；復發；機器學習；機器學習；決策曲線分析；重分類淨改善效益指數；臨床影響曲線

並列摘要

Background: Based on the International Agency for Research on Cancer (IARC) statistics in 2020, rectal cancer is the 7th most common cancer. Despite advances in treatment technology, recurrent events remain a critical clinical issue for rectal cancer survivors. The objective of this study was to identify important disease risk factors from the rectal cancer registry database, evaluate the prediction models developed for different factors, and analyze clinical characteristics. Methods: The analysis of 3,403 hospital-based cancer registry records was conducted between 2011 and 2020 from six cancer registries of hospitals. In this study, 16 independent variables were determined based on literature review and expert consultation. Predicted classifier models used were CART (Classification and Regression Tree), RF (Random Forest), C4.5 (C4.5 Decision tree), and C5.0 (C5.0 Decision tree). Regarding the problem of imbalanced sample categories, this study uses the extraction method synthetic minority oversampling technique (SMOTE) to solve this problem. The performance was assessed by accuracy, sensitivity, specificity, F1 score, kappa, Matthew’s correlation coefficient (MCC), and area under the curve (AUC). A clinical impact curve, decision curve analysis, and net reclassification improvement (NRI) were used to compare the clinical effectiveness of each prediction model. Results: The top three important risk factors for overall recurrence of rectal cancer were chemotherapy, combined stage group, and carcinoembryonic antigen (CEA) lab value. Based on the analysis of recurrence of patients with chemotherapy, the three most important risk factors were Carcinoembryonic antigen lab value, combine stage group, and surgical margins. It has been shown that C5.0 and C4.5 produce the best classification results for recurrence prediction, as well as an indication of their importance. C5.0 (0.8635) was found to be the best AUC value in all predictive models. C4.5 (0.8673) was the highest AUC value in patients with chemotherapy, according to the decision curve analysis, C5.0 had the highest predicted clinical benefit for the overall recurrence analysis. In addition, CART and C4.5 were predicted to have the highest clinical benefits for patients with chemotherapy. The results of the NRI analysis showed that only C5.0 had a significant in the overall recurrence analysis. Discussion and Conclusion: Recurrence may result in an increased risk of long-term survival due to the chemotherapy for rectal cancer. In order to improve cancer survivor management and surveillance, it is increasingly important to identify treatment-related factors that increase the risk of recurrence. These findings could be useful for developing effective prevention and surveillance programs for rectal cancer survivors at high risk during their follow-up. This class imbalance issue can be addressed in the future through a more advanced approach. It is one of the limitations of this study, as well as one of the potential directions for future research.

並列關鍵字

rectal cancer ； recurrence ； machine learning ； chemotherapy ； decision curve analysis ； net reclassification improvement ； clinical impact curve

參考文獻

Achilonu, O. J., Fabian, J., Bebington, B., Singh, E., Eijkemans, M. J. C., Musenge, E. (2021). Predicting Colorectal Cancer Recurrence and Patient Survival Using Supervised Machine Learning Approach: A South African Population-Based Study. Front Public Health, 9, 694306. https://doi.org/10.3389/fpubh.2021.694306

Google Scholar

Andrade, C. (2015). Understanding relative risk, odds ratio, and related terms: as simple as it can get. J Clin Psychiatry, 76(7), e857-861. https://doi.org/10.4088/JCP.15f10150

Google Scholar

Asoglu, O., Karanlik, H., Muslumanoglu, M., Igci, A., Emek, E., Ozmen, V., Kecer, M., Parlak, M., Kapran, Y. (2007). Prognostic and predictive factors after surgical treatment for locally recurrent rectal cancer: a single institute experience. Eur J Surg Oncol, 33(10), 1199-1206. https://doi.org/10.1016/j.ejso.2007.02.026

Google Scholar

Benson, A. B., Venook, A. P., Al-Hawary, M. M., Arain, M. A., Chen, Y. J., Ciombor, K. K., Cohen, S., Cooper, H. S., Deming, D., Garrido-Laguna, I., Grem, J. L., Gunn, A., Hoffe, S., Hubbard, J., Hunt, S., Kirilcuk, N., Krishnamurthi, S., Messersmith, W. A., Meyerhardt, J., . . . Gurski, L. A. (2020). NCCN Guidelines Insights: Rectal Cancer, Version 6.2020. J Natl Compr Canc Netw, 18(7), 806-815. https://doi.org/10.6004/jnccn.2020.0032

Google Scholar

Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5-32. https://doi.org/10.1023/A:1010933404324

Google Scholar

主題瀏覽